The Daily Byte
Breaking
Welcome to The Daily Byte — real stories from across Pakistan.
Saturday, 21 November 2026 · Pakistan

The Daily Byte

Stories from Pakistan that deserve to be heard
Minor World

Edge0-145B AI model runs on a laptop

By The Daily Byte ·14 Nov 2026 ·4 min read
A silver MacBook Pro laptop, the kind of high-memory laptop Edge0 says can run its 145B AI model locally

Edge0 has open-sourced two large language models, Edge0-145B and Edge0-120B, together with the inference framework and tools needed to run them on a personal computer. The company announced the release on 12 November, describing Edge0-145B as the first open-source on-device AI model with 145 billion parameters.

The claim matters because capable AI models have, until now, mostly lived in data centres. Running them locally has meant either accepting a much smaller model or paying for cloud access, with the fees, connectivity and data-handling questions that come with it. Edge0 says its new model brings coding, reasoning and agent work that rivals frontier cloud models onto a compatible laptop.

How a 145B model fits on a laptop

Edge0-145B is a mixture-of-experts model. Of its 145 billion parameters, only about 10 billion take part in computation on each step, because only a small set of specialist sub-networks, or experts, is used for any given token. What the company says is new is that groups of 64 experts share two large matrices, with each expert keeping only a small set of parameters of its own. Previously, every expert stored a complete set of matrices.

The model was also trained in its on-device form from the outset, rather than being compressed down from a cloud model afterwards. At run time, frequently used experts stay in GPU memory, an expert library of about 40GB sits in system memory and is loaded on demand, and about 48GB of vocabulary data stays on the SSD and is read as needed. The GPU and CPU work in parallel, and speculative decoding checks several candidate tokens at once.

Edge0 reports peak memory of 29GB on a compatible laptop, such as a MacBook Pro, with decoding above 80 tokens per second, prompt processing above 1,100 tokens per second and typical power draw of about 120 watts. The context window is 128,000 to 256,000 tokens. The company names laptops with 48GB of memory as target devices, along with workstations with 64GB. The weights are published on Hugging Face, and the framework and guides are on GitHub, with an Edge0 app for users who do not want to set anything up. The company says the model works offline, without an account and without usage limits.

Benchmarks that need independent testing

On Edge0's own figures, the larger model scores 64.6 on DeepSWE 1.1, ahead of Qwen3.8-Flash-Next at 58.7 and DeepSeek-V4-Flash-0731 at 54.4, and 86.52 on Terminal-Bench 2.1. It scores 90.4 on GPQA-Diamond, close to DeepSeek-V4-Flash-0731 at 90.8 and Claude Opus 4.6 (Max) at 91.3, and 80.6 on IFBench. Averaged across six benchmarks, Edge0 reports a mean of 72.8, against 71.7 for Qwen3.8-Flash-Next.

Those numbers come with an important caveat: the scores for Edge0, Qwen, DeepSeek and Claude come from Edge0's own internal evaluations, and the company says the full evaluation setup, hardware environment and reproduction instructions are being released with its technical report. There are no independent measurements yet, so the figures should be treated as the developer's claims, not as a settled verdict.

What the model is shown doing

Edge0 says the model can carry a task through from a single request, planning the steps, writing code, running it, fixing errors and checking the result with screenshots. Its demonstrations include building a 3D scene of Manhattan in Blender, creating a medieval city with more than a thousand homes and a Gothic cathedral, producing an explainer video about Tesla's Roadster, writing the control program for a drone and designing a two-layer circuit board.

The release follows Edge0-35B, a phone-scale model released in September that the company says runs on an iPhone 18 Pro Max using about 2.5GB of memory. If the laptop claims hold up in independent testing, the release would narrow the gap between what a personal computer and a data centre can do with AI, and let developers keep code and data on their own machines. Until that testing happens, it remains a free download and a promising experiment rather than a proven replacement for cloud models.